Single-Lead AI-Enabled ECG Serves as a Powerful Noninvasive Screening Tool: Study Confirms

Hyperkalemia is a frequently fatal electrolyte disorder particularly prevalent in patients with chronic kidney disease (CKD) and heart failure (HF), yet current detection methods remain hindered by the need for invasive venous blood sampling and clinical facility access. Although previous deep learning models using 12-lead ECG data have shown promise, a significant clinical gap exists for scalable home monitoring, leading Gongzheng Tang and colleagues from the Peking University Health Science Center to develop Pocket-K with the aim of providing a reliable, single-lead AI screening solution.

Therefore, the multicenter observational study utilized a dataset of 62,290 ECG–potassium pairs from 34,439 unique patients across two independent Chinese tertiary health systems between 2016 and 2024 to fine-tune a pretrained foundation model for detecting serum potassium levels greater than 5.5 mmol/L. By extracting Lead I data and excluding samples with laboratory-confirmed hemolysis or those lacking a paired measurement within a strict one-hour window, the researchers established primary and secondary endpoints focused on diagnostic discrimination across internal, temporal, and external validation settings.

Key Clinical Findings of Study Include:

• High Internal Accuracy: The study found that Pocket-K achieved a robust AUROC of 0.936 with a sensitivity of 83.33% during initial testing.

• Robust External Reliability: Even when tested in a geographically independent health system using different hardware platforms, the model maintained a commendable AUROC of 0.808, affirming the study's generalizability.

• Superior Severe Detection: For clinically urgent moderate-to-severe elevations of 6.0 mmol/L or higher, the AUROC increased to 0.861 in the external validation set, as the study noted.

• Reliable Rule-Out Capacity: The system demonstrated an NPV of 99.91% for moderate-to-severe events, suggesting the study's potential for identifying low-risk states.

• Rapid Clinical Inference: A handheld prototype from the study successfully executed near-real-time risk estimation via a smartphone application, returning risk probabilities for normokalemic and severe examples within seconds.

The results suggest that this foundation-model-based approach provides stable discrimination for hyperkalemia across heterogeneous populations, maintaining a high NPV above 99.3% while accurately tracking individual longitudinal potassium trajectories.

These findings indicate that clinicians can utilize single-lead AI systems as non-invasive rule-out tools to enhance the frequency of potassium surveillance for patients at high cardiorenal risk.

Although the study is limited by its retrospective design and geographic concentration in China, it underscores a clear need for future research using native consumer recordings to assess the impact of AI-guided screening on clinical decision-making.

Reference

Tang G, Zhao Q, Nie G, et al. Artificial intelligence-enabled single-lead ECG for non-invasive hyperkalemia detection: development, multicenter validation, and proof-of-concept deployment. 2026.